90 likes | 220 Vues
In Week 8 of the REU program at UCF, we fixed a critical bug in the SVM code, leading to substantial improvements in face/non-face recognition accuracy, rising from 75% to 95%. Binary SVMs were run across six classes, finding poles for emotions such as happiness, sadness, and anger. Database expansion techniques included shifting, rotating, and modifying images to generate diverse datasets—yielding 200 variations for each image. Initial testing results showed varying effectiveness of different SVM types, with plans to verify and refine approaches in upcoming experiments.
E N D
Adam Yeh UCF Computer Vision REU Week 8
Problems • Fixed bug in SVM code • Data had to be normalized for better results • Faces/non-faces improved from 75% to 95%
Finding Poles • Recap: • Run all binary SVMs (6 classes -> 15 binary SVMs) • For each class, find the intersection of non-SVs from binary SVMs • Progress: • Ran on training of about 50 images/class • Found poles on all but one class • Need metric to measure “closeness” to a pole
Finding Poles • Happy • Sad
Finding Poles • Anger • Surprise
Finding Poles • Disgust
Database Expansion • Database • Shift image left, right, up, down • Rotate image by small degree • Modify intensities • Blur images • Scale images up and down • Apply Gaussian filters in different regions • For each image, generates 200 images
Database Expansion • Testing: Anger vs Disgust • Radial Basis SVM: 73% vs 52% on testing • Linear Basis SVM: 54% vs 65% on testing • Problem • Previous algorithm will break down if binary classifications do not yield good results
Next Week • Verify binary SVM results with larger database • Experiment with eliminating certain transformations from database • Manual one vs. all binary SVM • Switch to CoreSVM • libSVM runs ~1hr per binary SVM • Debug code • Verify SVMs are more confused